AI Sustainability and Carbon Emissions

Overview

AI sustainability refers to the environmental impact of large-scale AI training and inference workloads, particularly carbon emissions and energy consumption. The rapid scaling of AI model complexity and deployment has created a significant and growing sustainability challenge:

• Rapid AI growth led to a 150% rise in CO2 emissions from major tech companies between 2020 and 2023 (source: ITU).
• Uncontrolled AI system growth could generate approximately 102.6 million metric tons of CO2-equivalent annually.
• GPT-5 is approximately 8.6× more energy-intensive per query than GPT-4, with each GPT-5 response consuming approximately 18 watt-hours.
• At 2.5 billion daily queries, GPT-5-scale deployments could require 45 GWh/day — enough to power approximately 1.5 million US households.

Efficient liquid cooling is positioned as a key mitigating technology: it reduces the energy needed to remove heat (improving PUE), and — when using low-carbon-footprint dielectric fluids like Perstorp's SYNMERSE — can dramatically reduce the product carbon footprint of the cooling infrastructure itself. By conscious raw material sourcing, Synmerse DC's carbon footprint can be reduced a further 92% beyond its already-low baseline, providing a path toward net-zero-aligned cooling infrastructure.